The False Promise of the Humanoid

Robots made in our image are technology's greatest red herring

General-purpose, humanoid robots – robots that look like humans and promise to do all the things that humans can do – are not a new concept. We can trace the humanoid at least as far back as 700 BCE, when the ancient Greek poet Hesiod described the story of Talos, a giant bronze man built by Hephaestus who protected the island of Crete from invaders. Nor were the Greeks the only people fascinated by the humanoid - an ancient Taoist text written by Chinese philosopher Lie Yukou in 300 BCE spoke of a humanoid automaton, and the earliest known references of the Jewish golem date back to 300 CE, among many other examples scattered throughout history.

In 2024, it seems like we are on the cusp of realizing a millennia-old dream of a functional robot made in our own form. Entrepreneurs and investors have flocked to the category, investing countless engineering hours and billions of dollars this year alone. Many believe that humanoids will soon flourish in the workplace, factory floor, assembly line, warehouse, hospital, and even at home. Many companies are seeking to build the humanoid themselves (e.g., Tesla Optimus, Figure, 1X Technologies, Agility Robotics, Boston Dynamics, Sanctuary AI, Unitree) while others are focused on the underlying foundation models that will power these machines (e.g., Physical Intelligence, Skild AI).

What was once perhaps an idle fantasy is now, at least on the surface, an inevitability. Believers in the general-purpose humanoid all share some version of the following thesis:

If our world has been designed by and for humans, then the human form will be the most effective means of interfacing with it.

I believe this thesis is deeply flawed and misunderstood. I do not believe the humanoid is an inevitability. In principle, a human form factor may be highly versatile, but in practice, non-humanoid robot form factors[1] are better. I would go a step further and argue that general-purpose humanoids will hold the robotics industry back and could delay the adoption of beneficial robotic automation in our society. To understand why this is the case, it's critical to understand the challenges that robot companies face in getting to production – and why humanoids uniquely struggle to overcome these challenges.

---

Why Robots Fail

"Most people think robotics is a hardware problem – but it's a software problem." So starts a recent post from Sequoia Capital, announcing their investment into Skild AI, a company building foundational models for general-purpose robotic intelligence. Reframing robotics as software has been a recent trend amongst entrepreneurs and investors. It makes sense why this is desirable. It's nice to imagine that a traditionally capital-intensive category can be "solved" with a positive FCF-margin solution (a pitch, no doubt, that is particularly enticing to a VC investment committee). And to be fair, there's at least some technological evidence supporting this optimism - with hardware costs declining and large foundational AI models (potentially) presenting a new paradigm for robotic intelligence.

Unfortunately, the software vs. hardware narrative is not just wrong – it completely misses the point. Hardware and software are challenges that robotics companies encounter but they are not the primary reasons why these companies fail. There is a third understated and underrated challenge, and it's the hardest to overcome. I call it the operationalization problem.

Operationalization is the process of taking a robot, which has hardware and software components, and getting it to work in a real customer environment. It means integrating the robot both upstream and downstream of some existing process and connecting the robot to all the systems that exist in the customer environment. It means incorporating the robot into human workflows and teaching the human how to work with the robot (and sometimes vice versa). It means grappling with a seemingly infinite number of edge cases that were not identified in the lab or initial POC. It can mean starting this process from scratch for every new customer. It means knowing how to communicate with the customer in a language they can understand.

Over the past year, I have had hundreds of conversations with roboticists. The conversations that have been the most compelling were with individuals who have already been in the trenches. To give a sense of the scope of the "operationalization problem," I've shared a small handful of quotes below (anonymized to protect identities and privileged information).

In short, when you ask roboticists to talk about their battle scars, they talk about operationalization – not hardware nor AI. This is a significant narrative violation, and it unfortunately catches too many first-time founders unaware.

So why is this relevant to this post? Well, once we accept that operationalization is the key challenge, it becomes clearer why humanoids are not the ideal form factor for the future of robotics.

---

Why Humanoids Fail Harder

Simply put, humanoids are particularly bad at operationalization, in addition to also struggling at the hardware and AI layers. Here's why:

• Humanoid hardware is overly complex and integrating complex hardware is challenging: If building a performant robot is a complex engineering problem, building a performant humanoid robot is an extremely complex engineering problem. Humanoids have more degrees of freedom, more balance and stability issues, higher mechanical intricacy from trying to mimic the human anatomy (joints, limbs, hands, etc.), and more complex actuator problems and limitations, among many other challenges. This introduces complexity to the hardware layer, an area where most robotics companies would prefer to avoid complexity. Furthermore, it exacerbates many separate operational issues that I’m about to detail further – cost, reliability, safety, time to market, error rates, etc.

• Humanoid hardware is unnecessarily costly: The cost of humanoids currently dwarfs that of their non-humanoid counterparts. Many believe that the cost of a humanoid can eventually fall to low- to mid-5-figures, which is cheaper than the cost of a human laborer (from a corporation's perspective) or cheaper than the cost of some consumer purchases (from the consumer's perspective). However, no company has so far demonstrated they can produce such a robot. Furthermore, training, integration, maintenance, and upgrade costs for humanoids remain unknown until more of these humanoids are deployed. Of course, these costs will come down in the long run - but so will the costs of non-humanoid robots. Humanoid companies will attempt to pass the higher costs down to their customers, who will struggle to justify the high price tag when they can just buy a more specialized and cheaper non-humanoid.

• Complex maintenance supply chain: The supply chain for new humanoid parts can be quite complex. The hardware components are delicate, and due to the nature of the tasks, they need to be replaced quite frequently (n.b., this is truer of humanoids because they are mobile vs. stationary). Furthermore, customers are typically bad at knowing how to replace parts, which can create significant ROI-destructive downtime. One famous general-purpose robot company told me that they had to fly out employees to customer locations weekly to fix and replace parts. For another famous humanoid robot company, their solution was simply to replace the robot instead of fixing certain parts (i.e., they literally just ate the cost of the components that did not need to be replaced).

• Humanoid companies are overhyping their capabilities: Humanoid robot companies are unnecessarily exaggerating their capabilities to a public audience. Insiders at several high-profile humanoid robot companies have confided that their demos range from misleading (e.g., cherry-picking 1 out of 100 attempts to make the robot appear more capable) to borderline fraudulent (e.g., saying the humanoid in the demo is standing on two legs, when in reality it is pillared to the ground and obscured by a table). While unfounded claims may serve to attract investment dollars, I believe they will also serve to disappoint customers who expect something quite different than what they will get – a dangerous double-edged sword. In some places, the damage is already done – I've spoken with several robotics R&D/innovation leaders across multiple end markets who roll their eyes when I mention the word humanoid.

• Humanoids are relatively unsafe: Humanoids are more prone to collapse than their non-humanoid counterparts. This is driven by physics – humanoids often have two points of contact with the ground (vs. three or four, which is passively stable), a high center of mass, and often carry payloads outside the cone of stability (n.b., the area in which a humanoid can sway without losing balance). They can also be quite heavy, further posing a safety risk to nearby humans. These environments can range from manufacturing settings populated by healthy adults to home environments filled with children, babies, and pets.

• Foundation models that can power humanoids are far from realization: While advances in ML/AI are occurring rapidly, the industry is far from having a useful foundation model. Companies like Physical Intelligence and Skild AI are doing interesting work in trying to recreate a GPT-like model for robotics. But this will take longer than most people think. First, there's a massive data shortage and no ideal or easy method of scaling data collection, unlike how the internet facilitated training LLMs. Second, the nature of the real world is far more open-ended and complex. Already hard problems, such as walking up different types of stairs, can become infinitely more challenging when you want robots to deal with other factors simultaneously, such as carrying objects of different weights and shapes. Third, if we pattern match the length of time it has taken to achieve useful computer vision models or LLMs, it's probably safe to assume at least a similar trajectory for robot models. And fourth, even if we were to have a robot model on par with the best LLMs today, it would be useless in my opinion, as the number of errors and hallucinations that are acceptable for language-based tasks is higher than what is acceptable for robotics tasks.

• Too many humanoids are solutions in search of problems: One piece of advice I consistently heard from robotics veterans was to never build a robot before you've identified a problem you want that robot to solve. I am not saying that robot companies can't have a more ambitious, long-term vision of becoming general-purpose. I also am not saying that all humanoid robot companies are making this mistake. What I am saying is that some - if not many - humanoid robot companies do make this mistake, and it’s because they believe the humanoid form factor will be so useful that commercialization will be easy. Spoiler: it won't be.

• Humanoids will take too long to reach market: Given the factors already outlined, we can conclude that humanoids will take longer to reach market than non-humanoids. Non-humanoids have fewer open-ended research problems and fewer hurdles to achieving production readiness. By the time these problems have been solved for humanoids, many non-humanoids will already be embedded in the most relevant end markets. Competition will be fiercer and true whitespace will be less abundant. Furthermore, robots are usually sticky products – so good luck ripping them out. Strategically speaking, it just makes more sense to build a non-humanoid that can achieve PMF quickly and then slowly embed more general-purpose capabilities into the robot over time.

• Humanoids will be left with immensely challenging tasks: By the time humanoids reach market, only the most challenging tasks will be available as whitespace. For less challenging tasks, humanoids will need to be able to displace existing automation solutions, which will either be handled by general-purpose non-humanoids or traditional robots. It’s worth noting that traditional automation providers have already automated as much as possible in most end environments today. Furthermore, due to rising labor costs, traditional providers are in fact expanding their capabilities because more expensive projects are becoming economically viable. This is all to say that humanoids will face exacerbated deployment challenges relative to other approaches.

• Humanoid mobility creates additional operationalization challenges: Humanoids trade efficiency for versatility. From a customer’s perspective, if the humanoid is not able to accomplish tasks faster than non-humanoid alternatives, then the humanoid must be able to make up for that loss in efficiency by being able to go handle other tasks that the non-humanoid cannot do during idle time. However, this introduces a new operationalization problem that has not yet been solved – how do you coordinate and herd a group of humanoid robots to accomplish a set of (probably) amorphous tasks (because the amorphous, weird tasks will be the only ones left untouched)? This isn’t impossible, but it is another layer of challenging operational complexity.

• Humanoids will struggle with reliability and ROI:  Given the factors already outlined, humanoids will struggle to be reliable and provide as much ROI as their non-humanoid counterparts. Failing on reliability and ROI has been a death knell for many robotics companies. This is the heart of the operationalization problem.

---

Where Does My Thesis Go Wrong

A lot of smart people will disagree with this post. I think they're wrong, but for fun I'll play devil's advocate.

On a surface level, you could take the opposite stance of anything I've said so far. For example, you could try to argue that humanoid hardware costs and complexity will be on par with that of non-humanoids very soon, or that we're only a couple years away from a GPT-like foundation model for robotics, or that the operationalization problem is trivial. But that's a bit boring and will sound fairly repetitive. Instead, I'd like to call out a few other arguments:

• Humanoids are psychologically appealing: Maybe it doesn't matter that the humanoid is a suboptimal form factor. The concept of a humanoid is undeniably captivating on a deep level with the human psyche – humanity has been talking about humanoids for thousands of years! It's possible that our desire to realize the humanoid will eclipse the barriers to their realization. Who knows...if all the smart people decide to go spin their wheels on humanoids, that might even give humanoids an advantage.

• Certain environments and end markets may be so chaotic that only a humanoid form factor is useful: I previously argued that in most end markets general-purpose non-humanoids will be able to commercialize more quickly than humanoids, even if the humanoid form factor is more versatile. But what if there are certain environments that are so chaotic that only a humanoid will ever be useful? I'm willing to entertain this. One potential example that comes to mind is the home environment, which is an enormous end market. It's possible that the non-humanoid wins in enterprise applications in the next ten years but the humanoid wins the consumer segment over the next twenty (though I would still bet against it). 

• Humanoid Convergence: It’s possible the non-humanoids that are successful in the near-term will eventually adopt more humanoid features, whereas the unsuccessful humanoids will need to become less humanoid in order to survive. In other words, over a longer time horizon, some of these form factors, which look quite different today, may converge.

---

Parting Thoughts: Hype -> Grind -> Graveyard

Normally, when you build a company, your trajectory should look like this:

In other words, you start your company, you grind really hard, and then either your company dies or it succeeds. If it succeeds, you earn the hype. This is “Grind -> Hype or Graveyard.” It’s the natural order of things.

Robot companies don't look like “Grind -> Hype or Graveyard.” Instead, they look like “Hype -> Grind -> Graveyard”:

For robot companies, hype is concentrated at the early-stage – typically at the Seed and Series A. During this phase, the robotics companies have a long-term vision, deliver an early prototype/MVP, and engage with a few design partners. Afterward, robotics companies enter an excruciatingly long period of "Grind." This is where the operationalization problem becomes apparent - customers fail to convert, integration costs remain high, the long-tail of errors appears unsolvable, and it's hard to manage burn. Inevitably, almost all these companies end up in the graveyard. The few historical success stories, frankly, aren’t that remarkable.

This is basically the opposite of how startups should work. It simply shouldn’t be the case that robotics companies are most attractive when they are all vision and no product, and least attractive after they build the product.

It's incredibly important that this changes. There are so many exciting problems that robots can eventually solve, particularly after embedding next-gen AI into their stack. However, “Hype -> Grind -> Graveyard” cannot change until the industry better understands why robot companies fail. And the humanoid is unfortunately not pushing the industry forward, but holding it back.

I'm investing significantly in technologies that I believe can transform the real world. If you're a founder and this post resonates with you, my DMs are open. If you hated this post, hit me up for a healthy debate.

---

[1] Not to be pedantic, but it's worth noting that the term humanoid is a spectrum - you could call a robot that looks virtually indistinguishable from a human a humanoid (e.g., Ava from the film Ex Machina), but you could also call a stationary arm a humanoid because, well, it has an arm. In this post, when I refer to humanoid, I'm generally referring to robots that are almost fully human in form (i.e., two arms, two legs, hands, feet, a torso, a head).